Decision Trace™

The Decision Trace: prove what your agents did.

Behavry is the independent system of record for autonomous agents. One signed account across every vendor they cross, verifiable by anyone you hand it to. The policy evaluated, the identity asserted, the authority delegated, the inputs scanned, generated inline by a control plane the agent cannot reach.

/001 · ANATOMYDECISION TRACE™

A log records an event. A Decision Trace records what actually happened.

/002 · WHY IT HOLDS UPFOUR PROPERTIES

Four properties a log can't claim.

01 · Independent

Signed by the control plane, not the agent.

02 · Deterministic

Same inputs, same decision, every time.

03 · Hash-chained

Linked to causal parents. Tamper-evident.

04 · Inline-only

Out-of-band tools cannot generate it.

/003 · RESOLUTIONSALLOW · DENY · INTERCEPT

Every decision resolves three ways.

Allow, Deny, Intercept. Each structurally distinct, each recorded as evidence. Enforcement is earned over time; the record is produced from day one.

Allow

Permit and proceed. Admissible under current policy. A Decision Trace is recorded as evidence.

Deny

Block before execution. The request never reaches the target. The agent receives a structured refusal it can reason about.

Intercept

Pause for human authority. Held for delegated authority the agent does not hold. An approver decides.

/004 · THE DOCTRINEBEHAVRY · RECORD

"The entity that acts cannot attest to its own behavior."

Agents cannot credibly attest to themselves. Neither can their vendors, nor the SIEM. A record kept inline, and structurally independent of the actor, is the one worth relying on. Behavry produces it.

We'll reconstruct a real cross-platform sequence and hand you the signed record

/005 · VERIFY IT YOURSELFNO ACCOUNT REQUIRED

Don't take our word for it. Verify one yourself.

A signed APR evidence package from a synthetic AI technician session: one allowed read, a password reset held for approval and then run, and two denials. Download it, drop it into the open verifier, and check the math yourself.

DOWNLOAD THE SAMPLE PACKAGE → NOW VERIFY IT YOURSELF →

Also grab the matching trust anchor · supplying it is what makes the result independent of Behavry. Package version 1.2; needs behavry-verify 0.14.0 or later. The AI technician boundary tests use the same sample.

/005B · HOW TO VERIFYPUBLIC SAMPLE · SYNTHETIC · ONE ACTION

How to verify an AI agent audit trail

One action, followed from request to verification. Every value below is copied from the public sample above, APR-2026-DEMO-000002: an AI IT technician working ticket T-4821 in a demo tenant, signed with a demo key. The package is synthetic and its five decisions were scripted; no target system was involved. Where the sample does not carry a field, this walkthrough says so instead of filling it in.

01 · The request

Ticket T-4821 asks for a password reset for j.rivera@acme.example. The agent's first reset for that user was held for approval (evt-demo-0001), and the identical call ran after approval (evt-demo-0002). The fourth event in timeline.json is the one followed here: the same tool, aimed at a different user.

FieldValue in the sample
idevt-demo-0003, in session sess-demo-it-4821
action · mcp_server · tool_nametool_call to msp-gateway-demo, tool cipp_reset_password
targetm365:acme.example/user/c.ng@acme.example, a user not named on the ticket
timestamp2026-11-13T14:04:00+00:00

02 · Agent identity and delegated authority

The event names the agent as agt-it-technician-demo in tenant tenant-demo-acme, and is signed by key behavry-demo-ed25519-2026-11. Its delegation_graph.json has two nodes: the requester j.rivera@acme.example, with ticket halo:T-4821 as context, and the agent beneath it. The event records the requester as requester_id j.rivera@acme.example, over channel psa, with requester_verified: false. So the sample shows on whose behalf the agent said it was acting, and states that this claim was caller-asserted, not verified. It carries no authorization records.

03 · The policy decision

FieldValue in the sample
policy_idticket_target_mismatch
policy_resultdeny
policy_reasonTarget is not the user named on ticket T-4821
enforcement_posture · execution_statusenforce · not_forwarded
event_hash88a94568…734989c5, hash version 9

The record says the call was denied in enforce mode and not forwarded to the target. It records which policy decided and why, but policy_bundle_digest is empty, so it cannot show which revision of the policy was in force.

04 · Verify it against the matching trust anchor

  1. Download the sample package and the matching trust anchor.
  2. Open the public verifier, add the package, and add the trust anchor file. Nothing you submit is stored. To check offline instead, the same open-source verifier runs from the command line; it needs behavry-verify 0.14.0 or later.
  3. Confirm that the report says the result was checked against the trust anchor you supplied. That line is what keeps Behavry out of the check.
$ behavry-verify --package apr-APR-2026-DEMO-000002.zip \
                 --trust-anchor behavry-trust-anchor-demo-2026-11.json
VERIFIED
✓ Signature valid               Ed25519 over the canonical manifest, key behavry-demo-ed25519-2026-11
✓ Signer fingerprint matches    sha256:e384a6966c311acf
✓ Merkle root matches           recomputed over 5 event hashes
✓ Event count verified          5 events
✓ Event signatures valid        5 of 5 events individually signed and verified
✓ File digests match            9 files match the digests in the signed manifest
✓ Event content matches hashes  5 decision event hash(es) recomputed from content
– Authorization records valid   package carries no authorization records

Then change a signed field in a copy. Set evt-demo-0003's policy_result from deny to allow in timeline.json and verify again:

NOT VERIFIED
✗ File digests match            timeline.json digest differs
✗ Event content matches hashes  content does not reproduce the signed event_hash for event(s) 'evt-demo-0003'

Output from behavry-verify 0.14.0 on November 18, 2026, abridged. The full report explains each skipped check.

What a pass establishes

The manifest was signed by the key in the anchor you supplied. Every file in the package is byte-for-byte what the signed manifest lists. Each decision event's hash rebuilds from its own readable fields, including the verdict, tool, target, arguments digest, requester and execution status. The event hashes, their count and their individual signatures all match. Together, those checks show that no event was added, removed, reordered or edited since the package was signed.

What it does not establish

That the password reset did or did not happen in Microsoft 365. not_forwarded is the record's claim. It does not prove the requester was who they said: the record says they were not verified. It does not show the decision was correct, since these were scripted. And it does not show completeness: actions on paths that were not governed are not in the record.

05 · Corroborate the outcome downstream

The signed record shows what the control plane decided. Whether the outcome matched comes from a record the target system keeps. For the denial in evt-demo-0003, that is the tenant's own directory audit log showing no password reset for c.ng@acme.example around 14:04. For the approved reset in evt-demo-0002, it is the matching reset entry for j.rivera@acme.example, made by the integration's service identity. The sample is synthetic, so it ships with no downstream record. In your own test, collect that record from a source you control and reconcile it with the trace.

The held call and the approved call share one arguments digest (f16f645d…), so a reviewer can confirm that what ran was the request that was held. The package carries no authorization record for the approval itself; the approver appears only in policy_reason text.

This walkthrough is the fifth test in How to Evaluate an AI Agent Governance Platform: Five Tests to Run Before You Buy. The guide covers the other four, from scoped permissions to human approval. The earlier sample, package version 1.0 with its own trust anchor, predates file and event-content binding: an edited verdict in it still verifies.

/006 · WHAT IT IS FORINCIDENT REVIEW · AUTHORITY REVIEW

Two reviews the record makes possible.

The same evidence answers the question after something goes wrong, and the question before you let an agent do more.

Incident review

Reconstruct a cross-system sequence step by step: which agent acted, on whose authority, what policy decided, and where the chain was stopped. Hand the signed package to whoever needs to check it.

Authority review

Read an agent's history before widening what it may do. Earned Autonomy builds on this: a recommendation from the evidence, then a human decision. The A1 to A2 assessment in shadow mode is planned.

What is captured

Actions that pass through a supported, configured execution path. Activity that bypasses those paths is not in the trace, and a trace does not establish the truth of an agent's stated intent.

What verification needs

A trust anchor: the public key you choose to rely on. Verification checks integrity and authenticity under that key. Supplying the anchor yourself, rather than fetching it from us, is what makes the check independent of Behavry.

/ONWARDBEHAVRY · RECORD
  • THE FOURTH A

    Authentication, authorization and accounting stop at the boundary. Agents act past it.

  • TRANSPARENCY

    What we publish about our own agents, including the one on this site.

  • COMPLIANCE

    Control mappings and evidence support, produced as a by-product of enforcement.

See a Decision Trace from your own traffic.

BOOK A DEMO →MSP or MSSP? Discuss an engagement →